Medical digital twins are increasingly presented as dynamic computational representations that integrate longitudinal clinical, imaging, biological, behavioural, and patient-generated data. Yet a central problem remains unresolved: how should an explanatory inference be translated into a clinically permissible intervention? This paper proposes a governed closed-loop architecture for digital twins in multiple sclerosis care. The framework separates patient representation, abductive hypothesis generation, intervention recommendation, authorisation, execution, and outcome-based revision. Its central proposition is that epistemic confidence does not constitute intervention authority. A strongly supported hypothesis may justify investigation or clinician review without authorising treatment. The architecture therefore implements action-specific governance gates for evidence integrity, temporal validity, competing hypotheses, causal support, eligibility, professional authority, consent, and safe learning. It also defines an architectural invariant prohibiting any path from hypothesis to clinical action that bypasses an authorised decision state. Multiple sclerosis is used as a demanding case because relapse-associated worsening, progression independent of relapse activity, imaging activity, disability, cognition, biomarkers, comorbidities, and patient-reported outcomes may evolve asynchronously. The framework further introduces a measurable oversight-quality profile to distinguish substantive human review from ceremonial approval, and a bounded emergency path that admits urgent action through pre-authorised delegation with mandatory retrospective ratification rather than through exemption from the invariant. A trustworthy medical digital twin should therefore be treated not as an autonomous predictive replica, but as a governed epistemic and operational system whose evidence, hypotheses, recommendations, decisions, actions, and consequences remain traceable, contestable, and subordinate to legitimate clinical authority.